Abstract

Large populations of simple robots can solve complex tasks, but controlling them is still a challenging problem, due to limited communication and computation power. In order to assemble objects, have shown that a human controller can solve such a task. Instead, we investigate how to learn the assembly of multiple objects with a single central controller. We propose splitting the assembly process in two sub-tasks -- generating a top-level assembly policy and learning an object movement policy. The assembly policy plans the trajectories for each object and the object movement policy controls the trajectory execution.The resulting system is able to solve assembly tasks with varying object shapes being assembled as shown in multiple simulation scenarios.